Fitting Neural Textures and PBR Material Maps with ES (No Backprop)

Researchers released an open-source C++ tool that compresses neural textures and PBR material maps using Evolution Strategies instead of traditional backpropagation, eliminating the need for automatic differentiation or machine learning frameworks. The method encodes images as a low-resolution latent texture plus a small MLP decoder, achieving compression results like 32.2 dB PSNR at 3.27 bits-per-pixel for a 128×128×8 configuration, and can train a four-layer material in about 150 seconds on a CPU. The approach requires only basic dependencies (stb_image libraries and OpenMP) and includes optional techniques like quantization-aware training and finite differences polishing.